feat: add native Groq provider

[JNK234]

Closes #51 (roadmap B1).

Stacked PR. Targets fix/provider-defaults (#71), not main, because both touch ProviderRegistrations.scala, models.yaml, and the provider counts in tests.txt. Merge #71 first and this retargets to main cleanly.

Why

Groq was reachable only through the OpenAI-compatible workaround — untested and undocumented as a supported path. Its free tier is fast and generous, which makes it the most useful zero-cost option for workshops and classrooms, where throughput matters more than model quality.

What

• GroqProvider.scala extending OpenAICompatibleProvider (90 lines, mirrors TogetherProvider)
• groq descriptor in ProviderRegistrations.scala, apiKeyPrefix = Some("gsk_")
• groq: section in models.yaml listing only currently live models
• groq added to ModelRegistry.FALLBACK_CONFIG — every other provider had an entry; omitting it would have left Groq validation broken if the YAML failed to load
• groq case in ReasoningModelDetector
• Docs in PROVIDER-GUIDE.md and SETUP.md

The default model choice matters here

The obvious picks were traps. Verified against Groq's deprecation page:

Model	Status
llama-3.1-8b-instant	shuts down 16 Aug 2026
llama-3.3-70b-versatile	shuts down 16 Aug 2026
qwen3-32b	decommissioned 17 Jul 2026
deepseek-r1-distill-llama-70b	decommissioned Oct 2025
qwen-qwq-32b	decommissioned Jul 2025

Both Llama models are still listed under "Production Models" with no banner. Either would have shipped a provider that broke within days. The default is openai/gpt-oss-20b — Groq's own recommended replacement for llama-3.1-8b-instant, 131K context.

Groq retires models aggressively. models.yaml will need periodic revisiting; noted in both docs. This is an argument for #64.

Verified vs. assumed

Verified from Groq documentation:

• max_tokens is deprecated in favour of max_completion_tokens; both still accepted, and the base class already sends the right one per path — no change needed
• reasoning_format accepts raw|parsed|hidden, default raw (thinking inline in <think> tags). Hence extractThinking keeps both paths rather than only reading message.reasoning.
• gpt-oss models ignore reasoning_format entirely and always populate message.reasoning — the parameter is a harmless no-op for the default model, correct for Qwen
• Unsupported OpenAI fields (logprobs, logit_bias, top_logprobs, presence_penalty, n>1) return 400. The extension sends none of them, so no filtering was required.

Assumed, flagged honestly:

• reasoning_effort default for gpt-oss is undocumented — the provider passes through only what the user sets rather than injecting a default
• message.reasoning appears in the reasoning docs but not the formal response schema; extraction is defensive and degrades to None
• No live API call was made — there is no Groq key in the build environment, so wire-format correctness is documentation-derived, not empirically confirmed. Worth one real call before release.

Tests

73 passing, 0 failed on the rebased branch (71 from #71, plus 2 new Groq blocks). Deterministic — the harness swaps in DeterministicTestProvider, so no key is required. The drift guard added in #71 now also validates Groq's default model.

Left out

• include_reasoning support — needs new config surface, beyond this issue
• requiresReasoningFormat entry — reserved for models that 400 without it (OpenAI o-series); Groq works without
• Groq in the interactive demo assets — those need a real key and are not on the CI path